The next frontier: Moving human fear conditioning research online

Fear conditioning is a significant area of research that has featured prominently among the topics published in Biological Psychology over the last 50 years. This work has greatly contributed to our understanding of human anxiety and stressor-related disorders. While mainly conducted in the laboratory, recently, there have been initial attempts to conduct fear conditioning experiments online, with around 10 studies published on the subject, primarily in the last two years. These studies have demonstrated the potential of online fear conditioning research, although challenges to ensure that this research meets the same methodological standards as in-person experimentation remain, despite recent progress. We expect that in the coming years new outcome measures will become available online including the measurement of eye-tracking, pupillometry and probe reaction time and that compliance monitoring will be improved. This exciting new approach opens new possibilities for large-scale data collection among hard-to-reach populations and has the potential to transform the future of fear conditioning research.


Introduction and background
Fear conditioning investigates the processes by which organisms learn to associate specific stimuli with aversive or feared outcomes (Bouton et al., 2020).Fear conditioning has been one of the most enduring topics of publication in Biological Psychology being present in the very first volume of the journal ( Öhman, 1974).Since this initial paper, Biological Psychology has regularly published work on human fear conditioning and is one of the go-to journals for researchers in the field.For instance, in the period between 2006 and 2020 between 5% and 10% of the papers submitted to Biological Psychology annually listed 'Learning and Conditioning' as one of the content categories addressed.However, rather than focussing on the past we intend to provide an overview of where we see the future of the fieldwhether and how it is possible to investigate human fear conditioning remotely and online rather than in the psychophysiology laboratory.
In fear conditioning experiments, participants are typically exposed to a neutral conditional stimulus (CS+), such as a shape or a face, which is followed by an aversive unconditional stimulus (US), like an electric shock or an unpleasant noise (Lonsdorf et al., 2017;Ney, Luck, et al., 2022).Through repeated pairings, the participant learns to associate the CS+ with the negative outcome, leading to the development of anticipatory responses to the CS+ as well as a shift in valence in that the CS+ becomes unpleasant or fear evoking.In human research, a second neutral conditional stimulus (CS-) that is never paired with the US is usually used to control for physiological responding due to non-associative processes which can also be interpreted as a safety cue.Fear conditioning research is used to understand the mechanisms underlying fear and anxiety in both humans and animals and has significant implications for understanding the processes underlying anxiety disorders and post-traumatic stress disorder (PTSD).By elucidating the mechanisms of fear learning and fear memory, researchers aim to develop more effective treatments and interventions for these conditions (Craske et al., 2014;Graham & Milad, 2011;Lipp, Waters, et al., 2020;Ney, Crombie, et al., 2022;Vervliet & Boddez, 2020;Zuj & Norrholm, 2019).
In human fear conditioning research, physiological and self-report outcome measures are commonly used to assess different aspects of the fear response (Lipp, 2006;Lonsdorf et al., 2017).Physiological outcome measures provide indices of human fear that are not under voluntary control to the same extent as are measures of self-report (Bach & Melinscak, 2020).These measures include heart rate, skin conductance, eyeblink electromyography to assess blink startle, respiratory rate, pupillometry, and cortical activity (Berg & Balaban, 1999;Dawson et al., 2016;Lipp et al., 1994;Ojala & Bach, 2020).Due to their largely involuntary nature, physiological responses are considered to be the best available outcome measure for indexing human fear conditioning (Lonsdorf et al., 2017).However, while physiological measures can indicate the presence and intensity of fear, they do not provide information about the participants' experience of fear or its cognitive aspects.For example, they may not differentiate between fear and other emotional states, such as anger or excitement (Blumenthal et al., 2005;Boucsein, 2012;Lipp, 2006;Lipp et al., 1994).
Self-report measures rely on individuals' experiences and perceptions of fear (Lonsdorf et al., 2017).These can be assessed through questionnaires, interviews, or rating scales.Here participants describe their individual levels of fear, anxiety, or distress, or their perception of the likelihood that a CS will be followed by the US (threat expectancy) or the perceived CS valence (e.g., CS pleasantness).Self-report measures capture the cognitive and emotional dimensions of fear, providing insights into an individual's conscious experience and appraisal of fear-inducing stimuli (Boddez et al., 2013;Lovibond & Shanks, 2002).They can provide information about the quality, intensity, and individual interpretation of fear.However, self-report measures are susceptible to response biases and social desirability effects.Participants' cultural background, personality traits, and mood state may also influence their self-reporting of fear (Boddez et al., 2013).Further, while physiological fear responses reliably reduce during extinction learning, self-reported CS+ valence frequently does not or at a slower rate (Lipp, Oughton et al., 2003, Lipp, Siddle et al., 2003;Luck & Lipp, 2015), suggesting that self-report and physiological outcome measures are complementary rather than equivalent.Moreover, verbal instructions and knowledge about the CS-US contingency strongly affect physiological measures (e.g., for skin conductance responses, blink startle see Hamm and Weike (2005); Purkis and Lipp (2001) but leave reports of CS valence relatively unaffected (Luck & Lipp, 2015) suggesting differences in the processes they reflect.For these reasons, fear conditioning studies rarely only use self-report measures and will usually include a combination of physiological and self-report outcomes (Lonsdorf et al., 2017).However, the recording of physiological responses usually requires the participant to come to the laboratory which limits access to larger and diverse samples that can be accessed in contemporary online-based research.
In this article, we review the existing literature on remote online fear conditioning, an emerging field that predominantly relies on measuring self-reports of fear and threat expectancy.We argue that, despite this limitation, online fear conditioning studies can provide a surprising amount of information that will permit innovative exploration of fear conditioning that would not be possible in laboratory conditions.Further, some physiological and behavioural outcome measures currently used in the laboratory may be incorporated into online fear conditioning in the near future, including eye-tracking, probe reaction times, and pupillometry, suggesting that online fear conditioning may become more comparable to laboratory-based research over the next decade.

Evaluative conditioning
It should be noted that the notion to assess emotional learning online is not entirely novel.Research on evaluative conditioning which assesses the acquisition of likes and dislikes (for reviews see De Houwer et al., 2001;Hofmann et al., 2010) has utilised remote testing for some time (see for instance Green et al., 2021;Heycke & Gawronski, 2020).Evaluative conditioning is procedurally similar to fear conditioning, in that a-priori neutral stimuli (CSs) are paired with a-priori pleasant or unpleasant stimuli (USs) with the aim to render them pleasant or unpleasant respectively.Evaluative conditioning frequently employs the picture-picture paradigm (Baeyens et al., 1989) which utilises pictures of neutral faces, cartoon aliens or geometrical symbols as CSs and emotional pictures, often drawn from the International Affective Picture System (IAPS, Lang et al., 2008), as USs.Evaluative conditioning is used widely in research on attitude formation and prejudice and has a rich history of theoretical debate both about the nature of evaluative conditioning itself and its relationship to Pavlovian fear learning (see Gawronski & Bodenhausen, 2006;De Houwer & Hughes, 2020).One difference between research on evaluative and fear conditioning is in the nature of the dependent measures that are employedwhereas fear conditioning traditionally relies primarily on physiological measures, evaluative conditioning relies mainly on self-report and the use of reaction time based implicit measures, implicit association tests or affective priming (Fazio & Olson, 2003), to assess emotional learning.There have been attempts to employ physiological measures in picture-picture paradigms, but the results are frequently not clear mainly because the US pictures that have been used were, in comparison to the electro-tactile shocks or loud scream USs used in fear conditioning, relatively low in arousal and did not support autonomic responding or blink startle facilitation reliably (see for instance Mallan et al., 2008).It would be interesting to revisit this question, however, using highly arousing USs such as pictures of mutilations or erotica.This more so as recent work in evaluative conditioning has highlighted parallels between the learning processes stimulated in both paradigms, such as the observation of return of differences in evaluation after extinction due to renewal or reinstatement (Luck & Lipp, 2020) or generalisation (Patterson et al., 2023) and there is evidence that prior evaluative conditioning will facilitate subsequent fear learning (Lipp, Luck, et al., 2020).Thus, research using evaluative conditioning paradigms can provide insights into emotional learning, both appetitive and aversive, that complement those derived from fear conditioning.

Online fear conditioning
The COVID-19 pandemic has greatly impacted laboratory-based psychological research, including fear conditioning.As a result, many researchers have turned to online platforms to continue their studies.This shift has allowed for increased accessibility and flexibility in participant recruitment and the ability to conduct experiments from remote locations, helping to mitigate the impact of the pandemic on research progress.Fear conditioning is one of many areas that has been adapted to online settings.However, although the original motivation of transferring fear conditioning online may have been due to the COVID-19 pandemic, recent studies have produced results that could not have been achieved using conventional laboratory-based experiments.Purves et al. (2019) first aimed to validate the use of a smartphone app (Fear Learning and Anxiety Response, FLARe, for details see McGregor et al., 2023) for the remote administration of a fear conditioning paradigm.The study involved 152 participants who completed the fear conditioning task, which involved acquisition, generalisation, and extinction phases and a contextual renewal phase 24 h following extinction.Different coloured circles were used as the CSs, while a loud female human scream was used as the US.Trial-by-trial US expectancy ratings were collected, and affective ratings for each stimulus were collected once per phase.The authors reported that the procedure successfully induced fear responses.Furthermore, the study found individual differences in fear conditioning, with participants who reported higher levels of anxiety exhibiting stronger fear responses to the conditioned stimuli.Finally, the results obtained from the smartphone app were similar to those obtained from a laboratory-based fear conditioning task, suggesting that the app is a valid tool for fear conditioning research (Purves et al., 2019).The use of this approach potentially allows for data collection from geographically diverse samples, which could help to enhance the ecological validity of research findings.Further, the app allows customisation of many of the settings and parametersfor example, FLARe would allow investigation of the effect of different CS types on fear conditioning, which is a historically important topic of research in human fear learning (Ney, O'Donohue, et al., 2022).McGregor et al. (2021) investigated the potential of the FLARe smartphone app as a tool for large-scale remote assessment of fear conditioning and its associations with anxiety.1146 participants completed the study, which consisted of a fear conditioning protocol as well as the completion of a Generalized Anxiety Disorder seven-item scale (Spitzer et al., 2006) and a self-reported history of anxiety diagnoses (Davies et al., 2022).Resembling results of laboratory-based studies, participants who reported higher anxiety levels exhibited higher threat expectancy ratings towards the CS-compared to participants with low anxiety levels.This work highlights the potential for online fear conditioning to be used as a tool for assessing anxiety and fear responses (McGregor et al., 2021).
This research group has also shown that it is possible to collect the large samples required for research on the role of genetics in fear conditioning.Purves et al. (2021) completed a large twin-study using FLARe to test the genetic and environmental factors that contribute to fear acquisition and extinction.This study reported the results from 1937 twins, including 538 twin pairs.Genetic factors accounted for 15% of the variance in fear acquisition, 30% of the variance in fear consolidation, and 15% of the variance in extinction learning.Further, a moderate proportion of concordance between genetic effects on fear learning and fear extinction was observed between participants.Interestingly, no evidence for an effect of common environment on fear conditioning emerged.This work demonstrates the potential for online experiments to make contributions to the fear conditioning literature through access to broader participant populations than achievable in the laboratory.
FLARe has also been used to test whether patients with anorexia nervosa (AN) show differences in fear conditioning compared to healthy controls (Lambert et al., 2021).This study recruited 64 women with AN and 60 healthy controls who completed all phases of the FLARe app protocol.Higher threat expectancy towards the CS+ and a larger difference between CS+ and CS-threat expectancies were reported for participants with AN compared to well controls.Further, participants with AN rated the CS-more negatively compared to well controls.Overall, the study provides preliminary evidence that women with AN show differences in fear conditioning compared to healthy controls.The study suggests the importance of considering emotional processing in the development of AN and the potential of fear conditioning as a biomarker for AN.
Although FLARe has been a success, it only supports a limited number of fear conditioning manipulations, which measure acquisition, generalisation, extinction, and contextual renewal.Cameron et al. (2022) tested the feasibility of using a novel online conditioning task to measure avoidance learning during the COVID-19 pandemic.119 participants were randomly assigned to either an experimental or a control group.The US was a loud female human scream presented concurrently with an image of a Caucasian female face depicting a fearful expression.
Participants were asked to wear headphones and turn them up to the highest setting and keep the headphones on for the duration of the task.Threat expectancy ratings were recorded after each trial and fear ratings towards each face were measured at the end of habituation, threat conditioning, avoidance shift (Engelhard et al., 2015) and test phases.In the avoidance learning phase, pressing the 'enter' key would always cancel the scheduled US when an avoidance cue was presented (an illuminated lightbulb in the top right corner of the screen).In the experimental group, an avoidance shift learning phase included the possibility of avoidance during the previously safe face (CS-), whereas the control group could avoid only during the CS+ .As the authors hypothesised, the option to engage in avoidance in the presence of the CS-led to elevated threat expectancy, demonstrating that remote online conditioning paradigms can be used to test avoidance learning for both threatening and safety cues (Engelhard et al., 2015;Cameron et al., 2022).
In a follow-up study, Cameron et al. (2023) tested whether counterconditioning with COVID-19-relevant stimuli reduced fear and persistent avoidance during the COVID-19 lockdown.The study involved 123 participants who were randomly assigned to either an experimental group that received online counterconditioning with COVID-19-relevant stimuli, or a control group that underwent standard extinction.Conditional stimuli were taken from the NimStim database (Tottenham et al., 2009) but were wearing COVID-19 related materials, such as masks.Counterconditioning involved pairing the CS+ with positive imagery intended to remind participants of pleasant living conditions prior to COVID-19.Participants in the counterconditioning group showed a significant reduction in threat expectancy and decreased avoidance following counterconditioning, relative to standard extinction.This finding highlights the potential of online counterconditioning as a possible intervention for reducing fear and persistent avoidance related to specific fears, such as those related to COVID-19 (Cameron et al., 2023).
One problem for online fear conditioning is the selection of an adequate US intensity and the assurance that the intensity will not be changed across the experiment.Berg et al. (2022) investigated the feasibility of individualized US calibration in an online fear conditioning study.The study involved 165 participants who were asked to calibrate the US by adjusting the volume of their headphones to the point of barely being able to hear a soft tone, which was determined to be the 0 dB volume intensity in the participants' audio setup.Next participants listened to a second beeping sound, at approximately 85 dB, and if this intensity was experienced as painful were asked to lower the volume to a level that the sound was very uncomfortable while still being bearable.The US was the sound of a fork scratching over a slate used in previous studies (Neumann & Waters, 2006) and was set at the same intensity as the second beeping sound had been set to.The authors found that the fear conditioning task with individualized stimulus calibration was feasible and effective, highlighting the potential of individualized stimulus calibration for improving the ecological validity of online fear conditioning tasks that take place without an instructor.However, when asked at the end of the experiment, one-third of participants self-reported lowering the volume after the calibration phase, but further self-reporting indicated that reduction of the volume did not render the US insufficiently aversive.
Further innovations have been made in other areas of online fear conditioning.Hauck et al. (2022) aimed to investigate how COVID-19-related anxiety affects fear learning and generalization in response to traumatic film clips.In this study, 220 participants completed a standard fear conditioning task using a 6 s film clip US from the film German Angst (Kosakowski, 2015).Studies using film clips as USs have begun to establish the validity of this approach in laboratory fear conditioning (Ney, Schenker, et al., 2022), but Hauck et al. is the first online replication of this approach.It was found that a higher level of COVID-19-related anxiety was associated with impaired fear learning and generalization, highlighting the role of individual differences in anxiety in the development of fear responses.However, only COVID-19 related anxiety was measured in this study and the authors were not able to conclude whether the effects were due to general rather than COVID-specific anxiety (Hauck et al., 2022).Regardless, this work provides proof-of-principle for the use of trauma film USs in online fear conditioning studies.
In an even more complex study design, Plog et al. (2023) aimed to replicate the role of phase synchronization and frequency specificity in the encoding of conditioned fear using a web-based fear conditioning task.The study involved 160 participants who were exposed to a series of Gabor patches, sinewave gratings that are seen through a Gaussian window and can be set to different spatial frequencies, that served as CSs, with one paired with a loud broadband white noise that served as the US.Across four experimental groups, the CS and US were modified by theta (4 Hz) or delta (1.7 Hz) frequencies, which were either in-phase or out-of-phase.Independent of frequency, phase synchronization significantly affected contingency awareness of the CS-US relationship but did not modulate ratings of CS valence or arousal.This study provides an example of a complex fear conditioning paradigm using remote, online experimentation, and suggests that declarative CS-US associations are facilitated by both theta and delta-frequency bands, largely replicating in-laboratory work done by this group (Plog et al., 2022).L.J. Ney et al.In summary, research has begun to show that not only is remote, online fear conditioning feasible, but that it is also useful for applications that are either complex or difficult to achieve in laboratory experiments.Over the next decade it is very likely that we will see a significant expansion of online fear conditioning research that will enable larger scale data collection of difficult to recruit populations.This will allow a better understanding of the mechanisms underlying both fear conditioning and psychiatric disorders.

Beyond self-report in online fear conditioning
One of the limitations of current online fear conditioning studies is the reliance on self-reports of US expectancy or CS evaluations as the sole dependent measures.This opens the results up to the effects of demand characteristics, i.e., participants providing responses that they believe commensurate with the outcome expected by the experimenters.Demand characteristics can be controlled for by measuring them explicitly in a separate sample (Cacioppo et al., 1992; see Luck & Lipp, 2015 for an example) or by the inclusion of measures that are less susceptible to demand characteristics.In laboratory-based studies this is achieved by the inclusion of physiological measures such as skin conductance.Recently, attempts have been made to employ measures that have shown validity in laboratory-based fear conditioning research online as well as exploiting some of the technical advances that have been made in online experimentation.

History of probe reaction time in laboratory fear conditioning
Probe reaction time has been used as an outcome measure in laboratory fear conditioning experiments since the 1980′s.Reaction time tasks in fear conditioning are employed as a secondary task, which builds upon the notion that cognitive resources to process information simultaneously are capacity limited (Dawson et al., 1982).In a secondary task procedure, performance of a primary task takes up most of the cognitive capacity, such that performance on a concurrent secondary task is impaired.For probe reaction time, deterioration in performance is indicated by a slower response such that slower probe reaction time is indicative of a more difficult, resource demanding primary task.Slower probe reaction times have been observed during a range of primary tasks, including word listening (Johnston & Heinz, 1981), rapid movement control (Ells, 1973) and letter-matching (Posner & Boies, 1971).Dawson and colleagues (1982) were the first to employ probe reaction time in fear conditioning.An auditory probe stimulus was presented during differential fear conditioning with visual CSs at six different stimulus onset asynchronies (SOA) relative to CS+ and CS-onset or offset.Probe reaction time was slower during the CS+ than during CS-, which was consistent across the six different SOAs, with the largest difference emerging at the SOA of 300 ms.These results were consistent with the notion that greater cognitive capacity was allocated to the processing of the CS+ than the CS-and rejected the notion that fear conditioning is a low-level associative learning process that does not require higher order cognitive processes.
Although this study provided an experimental foundation, some inconsistent findings started to emerge in later years.Inferring from findings by Dawson and colleagues (1982), one would expect slower probe reaction time during CS+ than during CS-, since CS+ requires more processing resources than CS-.This pattern of results was observed by others (Dirikx et al., 2009;Hermans et al., 2005), during fear generalisation using colour discrimination (Dunsmoor & LaBar, 2013) and fear conditioning with CSs differing in spatial orientation (Armony & Dolan, 2002) or when using movements in different directions as CSs in the context of pain learning (Karos et al., 2015).In contrast, faster probe reaction time during CS+ than during CS-was reported in one study (Björkstrand et al., 2022) and no difference was found in a visual discrimination task (Armony & Dolan, 2002).One potential explanation for the finding of faster probe reaction times during CS+ could be individual differences.Individual differences in secondary task performance were documented for participants who differed in the size of electrodermal responses to CS onset by Dawson et al. (1982).Responders with large electrodermal responses to CS onset tended to exhibit larger differential processing probe reaction times at an earlier SOA (i.e., around 300 ms) compared to those with smaller responses.This study also found that as soon as visual processing of the two CSs is finished (i.e., at around 500 ms) cognitive demands and differences in probe reaction time started to decline.However, a second peak of differential cognitive processing occurred at a SOA at the middle of the 8 s CSs (i.e., 3500 ms).On the other hand, for participants displaying smaller electrodermal responses to CS onset, the early processing demands at 300 ms and at 500 ms were roughly the same, since it took them longer to process the CSs or at least there was a larger variation in the early processing onset.In these participants, the later differential processing did not appear until 6500 ms after CS onset briefly before the US (Dawson et al., 1982).To further investigate individual differences, Barrett and Armony (2006) divided participants into high and low trait anxiety groups and found that high trait anxious participants had faster probe reaction times during CS+ than during CS-, whereas low trait anxious participants showed the reverse effect.Another interpretation of this phenomenon is the display of instant vigilance once exposed to an aversively valued stimulus after conditioning (Björkstrand et al., 2022).Similarly, it was argued that the presentation of an aversively conditioned stimulus might interfere with a concurrent irrelevant task and result in slower reaction times (Armony & Dolan, 2002;Williams et al., 1996).

Probe reaction time in online fear conditioning
Two studies have used reaction time as an outcome measure in online fear conditioning.Björkstrand et al. ( 2022) used a neutral auditory reaction time probe presented either 500 or 2500 ms during CS+ and CSthat lasted 5 s.They reported differences in probe reaction time during CS+ and CS-only for early probes and only during acquisition.Notably, these differences are opposite to the anticipated direction, as laboratory-based studies employing a similar protocol reported slower response times during the CS+ (Dawson et al., 1982;Dirikx et al., 2004;Dirikx et al., 2007;Hermans et al., 2005;Lipp et al., 1993).Björkstrand et al. concluded that their results were inconclusive regarding the utility of probe response times as a non-verbal learning index in remote fear conditioning experiments.
Our group also sought to determine the feasibility of using probe reaction times during online fear conditioning (Ney et al., 2023).Using a 2-day delayed extinction recall paradigm, we conducted two experiments, with the only difference being that Experiment 2 had nearly twice as many trials as Experiment 1.In contrast to Björkstrand et al. (2022), we found slower reaction times during CS+ (compared to CS-) during acquisition as well as during extinction recall.However, although both Experiments yielded slower probe response times during CS+ during extinction recall, probe response times differentiated between CS+ and CS-during acquisition only in Experiment 2. This appeared to be due to the increased number of experimental trials.Neither experiment yielded a difference between probe response times during CS+ and CS-during extinction.Notably, during the acquisition and extinction phases of Experiment 2 there were significant interactions between participant age and CS type, such that faster reaction times to probes during CS-were recorded in older participants.This finding corresponded to significantly lower self-reported threat expectancy during the CS-in older participants, suggesting that the older participants in this experiment were either more attentive or more readily able to learn the CS-no US contingency.Finally, our study used word categories as CSs, suggesting that human fear conditioning can be used to explore linguistic memory processes and that online delivery of fear conditioning is suitable to address a number of questions within experimental psychology.
Overall, our results suggest that the use of probe reaction times in remote online fear conditioning research is feasible, but effective use of this outcome measure may be constrained by boundary conditions, such as the number of trials presented and participant attentiveness.More research is needed to determine the conditions under which this measure can be useful in online conditioning experiments.Due to the current limitation of online experiments to self-report measures the prospect of being able to collect a non-verbal outcome measure remotely will expand the utility of online conditioning.

History of eye-tracking in laboratory fear conditioning
Eye gaze is a behavioural measure that has received relatively little attention in the fear conditioning literature.This is perhaps unsurprising since probe response time is a more common and accessible measure than eye gaze, and physiological measures such as skin conductance can help monitor a physiological aspect of emotional arousal as it reflects (sudomotor) sympathetic nervous system activity (Lipp, 2006).Yet, eye gaze measurement has some inherent advantages over other methods.It is a valid measure of attention, as experimenters can measure the speed and frequency with which an organism looks at a conditional stimulus instead of indirectly inferring attentional processing via probe reaction times.Eye tracking may also have potential as a measure of attention in online fear conditioning experiments, where other physiological measures such as skin conductance are not feasible.
The past decade has seen promising developments in the use of eye tracking to investigate how fear acquisition modulates attention to the CSs.In line with the notion that an organism's survival depends on its ability to quickly detect threat, several studies have demonstrated that the CS+ captures attention more strongly than the CS-.In circular visual search arrays with a neutral target and a CS distractor (the additional singleton task), participants make more frequent saccades to CS+ distractors than to CS-distractors (Hopkins et al., 2016;Koenig et al., 2021;Koenig et al., 2017;Mulckhuyse & Dalmaijer, 2016;Nissens et al., 2017), take longer to fixate on the target when the distractor is the CS+ (Mulckhuyse & Dalmaijer, 2016;Nissens et al., 2017), dwell longer on the CS+ distractor than on the CS-distractor (Koenig et al., 2017; but see Koenig et al., 2021), and exhibit shorter saccade latency when the target is the CS+ than when it is the CS- (Hopkins et al., 2016).However, while Nissens et al. (2017) found that the greater frequency of saccades to CS+ distractors relative to CS-distractors was more pronounced at earlier latencies, they and others (Mulckhuyse & Dalmaijer, 2016) have found no evidence that fear affects the latency of saccades to CS+ distractors.Collectively, this pattern of results has been taken as evidence that the CS+ distractors are not processed faster than other stimuli but acquire greater saliency within the oculomotor system (Mulckhuyse & Dalmaijer, 2016).Note that the impact of threat signals on attention is not necessarily straightforward; for example, Mulckhuyse et al. (2013) used an oculomotor selection task and found that shorter latency saccades were more strongly directed towards the distractor when it was the CS+ than when it was the CS-, while longer latency saccades were more strongly directed away from the distractor when it was the CS+ than when it was the CS-.
Eye tracking has also provided insight into how threat and safety signals affect divided attention.During fear acquisition itself, participants fixate longer on the CS+ than the CS- (Armstrong et al., 2022;Austin & Duka, 2010;Hopkins et al., 2015;Koenig et al., 2021;Koenig et al., 2017;Michalska et al., 2017;Xia et al., 2021) though see Reichenberger et al. (2020) although it is unclear whether such an attentional bias can be eliminated via extinction (Armstrong et al., 2022;Xia et al., 2021).Interestingly, Schmidt et al., (2015Schmidt et al., ( , 2017) used a cueing paradigm and found that the latency of saccades towards CS+ targets was shorter than for neutral targets, irrespective of whether the stimulus onset asynchrony between stimulus and cue was 50 ms or 1000 ms.However, the latency of saccades was shorter for CS-targets than for neutral targets only at longer stimulus onset asynchronies (Schmidt et al., 2017), in line with evidence that dwell time is longer for the CSthan for neutral stimuli during fear acquisition (Koenig et al., 2021; but not for faces Armstrong et al., 2022).This suggests that while threat and safety signals are prioritised due to their behavioural relevance, only threat signals capture attention.Eye tracking has also been used to assess attention in vicarious fear learning experiments, where participants view another person (the demonstrator) respond with distress to a CS+ and later exhibit fear when the CS+ is presented in the absence of the demonstrator and the US (Espinosa et al., 2020;Kleberg et al., 2015;Müllner-Huber et al., 2022;Skversky-Blocq et al., 2022).These studies have shown that participants fixate longer on the demonstrator's face during CS+ trials (Kleberg et al., 2015;Müllner-Huber et al., 2022;Skversky-Blocq et al., 2022).However, while longer fixations on the CS+ itself predict stronger skin conductance responses in the test phase (Kleberg et al., 2015), longer fixations on the demonstrator's face during CS+ trials are associated with reduced skin conductance responses in the test phase (Espinosa et al., 2020).In summary, there is collective evidence that threat signals modulate the allocation of attention in fear conditioning paradigms, but more work is needed to confirm the reliability of these findings given the heterogeneity of eye gaze measures reported across the studies.
Finally, there is promise that eye tracking could provide insights into fear learning and behaviour in more complex situations (e.g., when multiple stimuli, or multiple stimulus contingencies, are present).For example, eye tracking has been used in blocking and generalisation paradigms to investigate how attention is driven by interactions between the aversiveness and predictive value of a stimulus.It seems that exogenous and endogenous attention are driven by stimuli that have a higher aversiveness and/or predictive value (Barry et al., 2016;Eippert et al., 2012;Kampermann et al., 2019;Kausche & Schwabe, 2020;Koenig et al., 2017), but the aversiveness of a stimulus seems to have a stronger effect than its predictive value (Koenig et al., 2017;Wise et al., 2019).

Eye-tracking in online fear conditioning
Clearly, eye gaze can provide insights into how threat and safety signals modulate attention.However, despite the increasing technological and empirical support for eye tracking as a measure in internetdelivered experiments (Semmelmann & Weigelt, 2018;Yang & Krajbich, 2023), to our knowledge there are no published examples in the fear conditioning literature.Recently, our lab attempted to address this.We used webcam-based eye tracking (via WebGazer; Papoutsaki et al., 2017;Yang & Krajbich, 2023) in four online fear conditioning experiments, where each experiment utilised a different attentional paradigm (e.g., additional singleton task, dot probe) to assess attention to CS+ and CS-.Ultimately, our initial attempts were unsuccessful which was partly due to issues unrelated to eye tracking (e.g., we experienced stimulus timing issues with the jsPsych plugin; de Leeuw, 2015) and researchers should be aware of some of the challenges of implementing webcam-based eye tracking in online experiments.
Perhaps the most considerable challenge we faced was variability in the temporal resolution of the eye tracking (i.e., the sampling rate of participant's eye position through their web camera).Poor or variable temporal resolution is an obvious challenge for time-sensitive eye tracking measures such as the latency of saccades.There may be ways to overcome this issue, but these will likely come at the cost of reduced accuracy.For example, we calculated fixation dwell time as the proportion of eye tracking samples where gaze was directed towards a target, but we did not observe any convincing effects across any of our experiments (see Fig. 1).Fortunately, sampling rate was mostly variable between, rather than within, participants.Thus, we recommend that future studies use sampling rate as an inclusion criterion (i.e., exclude participants who have web camera setups with particularly low sample rates; Semmelmann & Weigelt, 2018).While it is disappointing that in contrast to previous laboratory-based studies (see Watson et al., 2019;Anderson & Britton, 2020) our initial experiment did not provide evidence for differential conditioning, the fact that we were able to monitor eye gaze remotely is encouraging.
The spatial resolution of webcam-based eye tracking also needs to be considered when designing an experiment.Previous studies have shown that WebGazer can estimate fixations within ~200 pixels on average (4-5 • of visual angle) and that this is sufficient to replicate in-laboratory gazing patterns (Papoutsaki et al., 2016;Semmelmann & Weigelt, 2018).It is unclear whether this was a meaningful limitation for our experiments, but it may be a consideration for some experimental paradigms.For example, finer spatial resolution is likely to be more important for an additional singleton task than a dot probe task.
Finally, a general limitation of online experiments is that the experimenter has little control over, or insight into, the participant's motivational and attentional state.This is particularly important if speed (in terms of saccades or a manual response) is an outcome measure.We found that many participants at some point during the experiment seemed to have eye gaze positions that were suggestive of directing attention off screen from the display.Given the duration of the experiment (minimum 30 min, maximum 1 h), participants may have become bored or distracted in their home environments.Being professional research participants (i.e., MTurk workers, Litman et al., 2017), they may have routine methods of working quickly through online studies with as little attention as possible (Chmielewski & Kucker, 2019;Saravanos et al., 2021) or of doing several tasks at once across different screens.Although it may be desirable to design fear conditioning experiments that are brief to avoid loss of participant attention, typically eye tracking measures will require many trials, which may make this solution not feasible.
We suggest that, when designing such experiments, researchers should develop ways to monitor or exclude participants who exhibit unsatisfactory behaviour.Examples include fixation controls (e.g., a trial only starts if a participant has been fixating for 500 ms over a twosecond window); multiple calibration and validation checks throughout the experiment; catch trials; and performance-based exclusions (e.g., participants need to maintain a certain level of speed and accuracy).Researchers should be aware that implementing such controls, while important, will increase the exclusion rate, which is already high in online experiments, and lengthen the experiment.For example, each of our experiments was accessed by more than 200 participants, and just 20-25% of these participants passed the initial calibration and completed the whole experiment (although other studies have had a better inclusion rate : Semmelmann & Weigelt, 2018;Yang & Krajbich, 2023).
Eye tracking has yielded several novel findings within the fear conditioning literature offering access to attentional processing of the conditional stimuli.While we have detailed some of the challenges in implementing eye tracking in online fear conditioning experiments, the broader literature is optimistic: several studies have already demonstrated the feasibility of webcam-based eye tracking in experimental psychology (Papoutsaki et al., 2016;Semmelmann & Weigelt, 2018;Xu et al., 2015;Yang & Krajbich, 2023).Future studies may be able to improve upon our efforts and conduct successful fear conditioning experiments using eye-tracking online, especially with rapidly improving technology and improved computer specifications available to average users.

History of pupillometry in laboratory fear conditioning
Pupillometry is the study of changes in pupil diameter in response to various stimuli and physiological states.The measurement and analysis of pupillary responses provide valuable insights into cognitive and emotional processes, neurological functioning, and autonomic nervous system activity (Mathôt, 2018), and is sensitive enough as a measure to be used in clinical diagnosis of certain neurological disorders (Chen et al., 2011;Ong et al., 2019).The pupil regulates the amount of light entering the eye and is therefore mostly responsive to changes in luminance (Beatty & Lucero-Wagoner, 2000;Spector, 1990).However, pupil size can also either enlarge (dilation) or reduce (constriction) based on cognitive and psychological processes (van der Wel & van Steenbergen, 2018).Pupil constriction and dilation are regulated by distinct neural pathways, both of which are well characterised (Mathôt, 2018;McDougal & Gamlin, 2008;Samuels & Szabadi, 2008;Szabadi, 2012;Wang & Munoz, 2015).Constriction is primarily controlled by the parasympathetic nervous system, with preganglionic fibres originating from the Edinger-Westphal nucleus and synapsing in the ciliary ganglion.Postganglionic fibres then innervate the iris sphincter muscle, causing its contraction and subsequent pupil constriction.On the other hand, dilation is mainly governed by the sympathetic nervous system.Preganglionic fibres arise from the intermediolateral cell column in the spinal cord, ascend to the superior cervical ganglion, and then postganglionic fibres reach the dilator muscle of the iris, leading to its relaxation and pupil dilation (Mathôt, 2018).
These neural pathways are at least partially responsive to psychological states and processes (Samuels & Szabadi, 2008;Szabadi, 2012).For instance, the parasympathetic pathway responsible for pupil constriction is influenced by cognitive effort and attention.When individuals engage in demanding cognitive tasks or direct their attention towards specific stimuli, pupil constriction tends to occur.This is believed to reflect increased mental effort or focused attention.Conversely, reduced pupil constriction may indicate cognitive load or reduced attentional focus.On the other hand, the sympathetic pathway involved in pupil dilation is closely tied to emotional arousal and the fight-or-flight response.When individuals experience heightened emotional states, such as fear, excitement, or arousal, pupil dilation often occurs.This response is thought to be linked to the release of norepinephrine, which facilitates dilation and prepares the body for action.Therefore, pupillometry has been used to provide valuable insight into psychological processes (Samuels & Szabadi, 2008;van der Wel & van Steenbergen, 2018;Wang & Munoz, 2015).
Pupil size changes as a function of fear conditioning (Borrego & Gardner, 1986;Finke et al., 2021;Jennings et al., 1978;Reinhard & Lachnit, 2002;Reinhard et al., 2006), with larger pupil constriction during the CS+ relative to the CS-generally observed (Finke et al., 2021).A recent meta-analysis has suggested that conditioning-induced changes to pupil size are relatively robust across studies.However, Fig. 1.Fixation hit rate to each CS during dot probe trials that were presented throughout fear acquisition (from Experiment 1 of an unpublished study from our lab, N = 32).Fixation hit rate was defined as the proportion of eye tracking samples where the participant was fixating on the target (within a region of interest of 200 pixels).There were no significant effects of time or CS and no interaction between these factors.
given the substantial differences in experimental design and the demands to control potential confounds that affect pupil size, high variability in the reported results is likely (Finke et al., 2021;Mathôt & Vilotijević, 2022).Changes in pupil size also seems to align with predictions made by formal learning models of fear conditioning (Koenig et al., 2018;Tzovara et al., 2018), further suggesting that they are valid markers of the learning processes tapped by this paradigm.
In principle, the evaluation of pupil size changes during fear conditioning offers unique advantages compared to other psychophysiological measures.Firstly, pupillometry provides a continuous and highly precise measurement with superior temporal resolution compared to skin conductance responses, which tend to have a one-second delay between stimulus onset and a response being emitted (Boucsein, 2012).Similarly, pupillometry avoids the notable weakness of other measures of fear conditioning (fear potentiated startle, probe reaction time) in that it does not require a probe stimulus, potentially putting it on par with the most sensitive measures of fear currently available, skin conductance, at least in terms of ability to track differential responses to CS+ and CS-(Jentsch et al., 2020;Korn et al., 2017;Leuchs et al., 2019;Zimmermann & Bach, 2020).Finally, measurement of pupil diameter does not require the attachment of electrodes to the participant or the purchase of expensive equipment but can be performed with consumer grade web cameras placed above or below the computer screen used to present the conditional stimuli.Interestingly, and in contrast to other physiological measures, it has also been reported that pupil responses do not habituate during fear acquisition (Leuchs et al., 2019), and evidence for conditioned fear emerges more rapidly in pupil diameter compared to other measures (Finke et al., 2021).Pupil responses also seem to occur in both early and late latency windows (Finke et al., 2021), which is similar to skin conductance responding which occurs across multiple latency intervals (Luck & Lipp, 2016;Prokasy & Ebel, 1967).Further, unlike skin conductance, pupil size change during fear conditioning seems to occur irrespective of contingency awareness, though there is some evidence that different processes may be involved in aware and non-aware participants (Finke et al., 2023).
Despite these potential benefits, the use of pupillometry in conditioning research is still evolving, with little consensus on data interpretation, optimal design requirements, data pre-processing, and response quantification.However, some response quantification techniques are beginning to emerge that may help to standardise fear conditioned pupillometry (Korn & Bach, 2016;Korn et al., 2017).Several practical challenges need to be considered, including the susceptibility of the measurement to potential confounding effects of physiological artifacts such as blinking and changes in physical stimulus parameters like contrast and luminance (Niehorster et al., 2022).In many cases, the requirement for a high-quality eye-tracker could present a financial burden to a fear conditioning laboratory, especially given that today's focus is on traditional measures such as skin conductance and fear potentiated startle, but the use of consumer grade web cameras may provide a lower cost alternative.Finally, there is little known about the neural and psychobiological mechanisms that mediate the effects of fear conditioning on pupillometry which makes interpretation of these data challenging (Finke et al., 2021;Ojala & Bach, 2020; although some research has begun in this area, see Leuchs et al., 2017).These factors have resulted in a slow uptake of pupillometry as a measure in laboratory-based fear conditioning, despite the potential advantages that this measure offers.

Pupillometry in online fear conditioning
We recently commenced the development of an application that will permit the measurement of pupillometry online.To achieve this, we converted mEyea free software conducting neural-network-based pupillometry (Mazziotti et al., 2021) into a JsPsych (de Leeuw, 2015) tool (Js-mEye) so that it can be used by researchers to collect real-time pupillometric data from participants' webcams (manuscript in preparation).Js-mEye was developed from mEye, which has been shown to detect changes in pupil location and size comparable to existing state-of-the-art technology (Mazziotti et al., 2021).Js-mEye can be set up almost entirely automatically by the participant in an online study in approximately 30 s with minimal manual calibration required.This process involves participants resizing a square (the region of interest [ROI]) over their eye.Js-mEye defines the pupil as the darkest region within the ROI.During automatic calibration, the threshold of darkness that defines the pupil is constantly adjusted.An algorithm then calculates the threshold predicted to identify the pupil most accurately and configures Js-mEye accordingly.After calibration, Js-mEye runs invisibly and monitors for stimuli added, removed, or changed by other JsPsych tools.When these events occur, Js-mEye logs pupil area (in pixels), the probability that the participant is blinking, a timestamp, and the exact document object model element(s) related to the event (manuscript in preparation).Like other tools, JsPsych manages Js-mEye's data (de Leeuw, 2015).As a limitation, participants are required to keep as still as possible throughout studies employing Js-mEye.
In the coming year, we will conduct a pilot study wherein Js-mEye's pupillometric capabilities are tested in a simulated online environment and compared to data collected with an EyeLink-1000 system in a more traditional laboratory-based setup to validate our methodology.To our knowledge, studies have not yet validated the utility of software to measure changes in pupil size in online samples, however we would like to note that tools other than mEye exist (Zandi et al., 2021).If successful, Js-mEye will enable researchers to quantify stimulus-induced changes in pupil diameter in larger and more regionally diverse samples than currently possible.Moreover, Js-mEye's public availability may reduce the cost of research conducting pupillometry in the laboratory and online which currently requires laboratories to have specialised, commercial equipment.

Future directions
As reviewed above, the emergence of online fear conditioning paradigms opens the possibility of novel experiments that may improve our knowledge about human fear.The chief advantage are the opportunities to assess large and diverse samples that go beyond the sizes and usual makeup of our routine laboratory-based samples.However, there are still many challenges with the implementation of online fear conditioning that need to be addressed.
One obvious question that arises is whether what is assessed in online experiments is indeed fear in a manner that would be comparable with the fear that is assessed in a fear learning experiment conducted in a laboratory setting.After all, the gold standard physiological measures that are recorded in laboratory experiments, skin conductance, heart rate changes or fear potentiated startle, are no available online and the non-self-report measures that can be recorded online are not selectively sensitive to fear. 1 Probe response time indexes cognitive resource allocation (Dawson et al., 1982), eye gaze indexes the locus of attention (Watson et al., 2019) and pupil diameter is sensitive to cognitive load, emotional valence or surprise (Beatty & Lucero-Wagoner, 2000).These processes are involved in fear learning, but not selective to it and would be observed during for instance non-aversive learning as well.However, the same limitation applies to skin conductance, heart rate changes or potentiated startle.A conditional stimulus will elicit skin conductance responses or a tri-phasic heart rate response regardless of whether the un-conditional stimulus it is paired with is an aversive electro-tactile stimulus or the non-aversive go-signal for a reaction time task (Lipp & Vaitl, 1988, 1990).Even differential blink startle potentiation has been observed during non-aversive conditioning (Lipp et al., 2003b), but this may be limited to differential conditioning procedures and not seen in 1 We would like to thank the anonymous reviewer who raised this important issue.
L.J. Ney et al. single cue conditioning (Lipp et al., 1994).Thus, observing changes in skin conductance, heart rate or potentiated startle does not necessarily ensure that the conditional response we are observing is fear and the process fear learning.
A second, somewhat related, question is whether online fear conditioning experiments can provide us with information that is relevant for our understanding of the acquisition, maintenance, reduction and return of human fear and anxiety in a manner that matches the contribution of laboratory-based fear learning research (see also Spix et al., 2021).The evidence to date which suggests that basic phenomena such as acquisition, generalisation, extinction or return of fear due to renewal as well as individual differences in fear learning can be observed in online experiments seems to support this claim (Purves et al., 2019(Purves et al., , 2021;;McGregor et al., 2021).Whether there are differences in what can be observed online and in the laboratory is an open questionone that we would be interested to address and that in itself can provide us with important information about human (fear-) learning.
Most of the procedural problems surrounding online fear conditioning research stem from the lack of control over confounding conditions that may affect outcomes and arise from conducting research in uncontrolled environments.Firstly, in the laboratory, most researchers use standardised calibration procedures that allow participants to select an unconditional stimulus intensity that is "unpleasant, but not painful" (Lonsdorf et al., 2017).In a remote online environment, the options for researchers to ensure this selection and monitor participant compliance are limited.FLARethe only software that is available on a smartphone implements a procedure that allows participants to calibrate the unconditional stimulus to a level that they find unpleasant.Moreover, FLARe will alert participants and record relevant data when headphones are disconnected or the volume is turned down during the experiment (McGregor et al., 2022).While some computer-administered experiments have implemented calibration procedures for the volume of the unconditional stimulus (Berg et al., 2022;Björkstrand et al., 2022), a procedure for recording participant compliance throughout the experiment to the extent implemented in FLARe has not been reported.Currently, computer-based experiments can use compliance checkssuch as surprise questions presented at a low volume or asking participants at the end of the experiment to self-report compliancehowever, these may not provide a valid measure of participant compliance with unconditional stimulus volume.
Relatedly, FLARe requires that participants use headphones, whereas computer-based online experiments generally permit speakers or headphones to be used.This issue has been acknowledged in the literature and it is generally agreed that headphone use should be compulsory for participants to ensure consistency across participants and to maximise the perceived aversiveness of the unconditional stimulus (Berg et al., 2022;McGregor et al., 2022).However, no recommendation for how this might be achieved in computer-based experiments has been provided to date.
Participant engagement is a second major issue in remote online fear conditioning.In laboratory-based experiments, it is possible for the experimenter to actively monitor participant engagement with an experiment, and participants are likely to feel compelled to maintain active engagement with the tasks provided as they know that they are being monitored.In online fear conditioning engagement checks are essential to test whether participants are attending to the experiment.However, it is critical for researchers to recognise that these checks may not necessarily index participant engagement with the task beyond them being at the keyboard while the experiment is running.Our most recent studydescribed earlier in this reviewfound that participants' eye gaze throughout the experiments frequently appeared to be off the screen once any trial-specific tasks were completed.This suggests thateven if providing apparently compliant dataparticipants doing remote online fear conditioning experiments may be only partially engaged in the task.Future experiments will need to develop methods for ensuring participant engagement, such as more interesting experimental designs, shorter experiments, or more comprehensive attention and engagement checks.
Finally, online fear conditioning experiments are still limited in recording responses other than self-report reliably and are almost entirely reliant on self-ratings of distress, CS valence, and unconditional stimulus expectancy.However, research by us and others has begun to explore the translation of methods that have proven sensitive for the assessment of fear conditioning in the laboratory into online environments.In particular, we believe that the use of probe reaction time, tracking of eye gaze and pupillometry are promising.However, the review provided here also illustrates the challenges to this approach highlighting that more research is needed to develop innovative methods for measuring behavioural or physiological responses in online conditioning experiments before they can become a real alternative for in-person laboratory procedures.Given the developments seen in the last few years that led to more sophisticated tools, greater availability of online services and more sophisticated computer equipment in private households, we are optimistic about the feasibility of the online approach to fear conditioning.

Conclusion
Fear conditioning is an important area of research that has played a critical role in creating knowledge about human anxiety and stressorrelated disorders.Fear conditioning has begun to move online and, at the time of writing, at least 10 studies have been published on the topic, most within the past 2 years.These studies have shown that online fear conditioning has the potential to generate significant new knowledge, though considerable challenges still exist.Future work that explores the opportunities and limits of remote fear conditioning will increase the resemblance between this mode of research and the conventional inlaboratory mode, which currently has distinct advantages such as access to psychophysiological outcome measures and active participant compliance and engagement monitoring.However, fear conditioning researchers are beginning to address these problems, and over the next 10 years it is possible that online fear conditioning research will provide data of the same quality as does in-person laboratory-based experimentation.This is an exciting prospect that will open new opportunities for large-scale data collection in hard-to-reach populations and may transform the way fear conditioning research is conducted into the future.

Funding
This work was supported by the Australian Research Council (DP180100869) and the National Health and Medical Research Council (APP1156490).

Declaration of Generative AI and AI-assisted technologies in the writing process
AI-assisted technologies (ChatGPT) were used to improve readability and language of initial drafts of the manuscript which subsequently were re-written and underwent extensive revision.

Declaration of Competing Interest
None.